AWH GitHub
PORTFOLIO / MMXXVI

Alexander Wessblad Hårdh

Systems & infrastructure builder working where code meets hardware: machine learning, networking, and embedded security.

Built at fifteen, from Nacka, Sweden.

§01 · Direction

Frontier capability, forced onto real hardware.

My long-term focus is physical AI: getting frontier-scale capability onto robots, humanoids and embedded systems through aggressive model compression, not ever-larger compute.

It's the direction I already work in. Four-bit QLoRA fine-tuning of a 7 B model, on-device inference on the Apple Neural Engine, and a compute cluster I built and operate myself. The next moves: robotics, hardware, and the harder open questions in affective and emotional modelling.

18
servers, 42U rack
133k
legal docs, preprocessed
4-bit
on-device inference
0→5
citation accuracy fixed

42U · 18 servers · 384 switch ports

§02 · Selected Work

Six systems, built end to end, from hardware to model to interface.

  1. A 42U rack built almost entirely from cold-outreach hardware donations. Three of 34 companies said yes, including Conapto, GleSYS and Rohde & Schwarz. Eighteen Dell, HP and Supermicro servers, 384 managed switch ports, and a Proxmox hyperconverged cluster, kernel-tuned to the metal with 1 GB hugepages and NUMA-aware scheduling. Currently standing up a private 4G LTE network on Ericsson RBS 6000 with an Open5GS core.

    • Proxmox VE
    • Linux
    • Server Hardware
    • RF Instrumentation
    • Networking
  2. A Swedish and EU legal LLM, built from the data up. I collected and preprocessed ~133,000 statutes, bills and court rulings (~9.9 GB) into a clean corpus from scratch, then ran a QLoRA fine-tune on a 4-bit base. It learned Swedish legal register and citation structure. When consumer-hardware limits degraded generation, I diagnosed it to root cause and proved the fix in a public PoC: a live RAG lookup layer at inference took citation hallucinations 0/5 → 5/5 on a five-example test, each answer verified against riksdagen.se — a proof of concept, not a general accuracy benchmark.

    • Mistral 7B
    • QLoRA
    • HuggingFace
    • LLM
    • NLP
  3. A real-time control-room simulator with selectable PWR, BWR and RBMK reactor models: point-kinetics, a multi-node thermal model, and xenon/iodine poisoning dynamics. The secondary loop uses IAPWS-IF97 water/steam properties, with numerical integration tuned to run on modest hardware and a scored scenario mode.

    • Python
    • Numerical Integration
    • IAPWS-IF97
    • Physical Modelling
  4. A privacy-first macOS control system that runs entirely on-device via the Apple Neural Engine and Core ML, mapping hand gestures and gaze to system actions. Four subsystems: hand-landmark tracking, gesture recognition, face/eye tracking, and action execution. Current focus is migrating hand tracking to YOLOv10 / HaGRID with Core ML export for full ANE utilisation.

    • Apple Neural Engine
    • Core ML
    • YOLOv10
    • Computer Vision
    • On-Device
  5. Voltage-profiled a STIHL iMOW 522c PCB with an oscilloscope to identify and verify active UART pads, establishing a serial debug interface into its embedded Linux system. From that foothold, working toward an unauthenticated root shell via bootloader and firmware analysis.

    • UART
    • Oscilloscope
    • PCB Analysis
    • Reverse Engineering
    • Embedded Linux
  6. A shipped macOS app that runs Mistral inference locally via MLX, generating a daily AI-written briefing from live news, Nord Pool electricity-price and weather feeds. The model runs on-device, with no cloud LLM API.

    • SwiftUI
    • MLX
    • Mistral
    • REST API

UART · 802.11 passive · RF

§03 · Field Work

Security, in the wild and under pressure.

Swedish Security Championship

National Final · 2026 · 14th of 22

Placed 14th of 22 finalists at the national CTF final — a two-person team where three were allowed, among the youngest present, in a field averaging 3–4 years older. Solved challenges across web exploitation, forensics, binary exploitation and reverse engineering. Halted the organiser's mandatory screen recording on-site by presenting GDPR-based legal arguments for participant privacy.

  • Web
  • Pwn
  • Forensics
  • RE
  • GDPR

Passive WiFi Security Analysis

Independent Research · 2025

Uncovered ARP storms, VLAN leakage and an exposed municipality-wide L2 broadcast domain through passive 802.11 capture and analysis in Wireshark, with no active probing and no packets sent.

  • Wireshark
  • 802.11
  • VLAN
  • ARP
  • Passive

§04 · On the Record

A device ban, formally declared void.

As sole complainant, I filed and argued independently against a school-wide device ban at Boo Gårds School, and had the decision formally declared void. The argument ran on mandatory student consultation (Education Act ch. 4–5), ejusdem generis, proportionality under the Instrument of Government, and an internal BYOD-network contradiction.

I then used the case to identify that Government Bill 2025/26:193 removes student consultation rights ahead of autumn 2026, then published the analysis to the broader student-rights community.

April 2026 · Nacka

§05 · Capabilities

Applied across projects, referenced to documentation.

Infrastructure & Networking

  • Proxmox VE
  • Linux · Debian / Ubuntu
  • 802.11 / WiFi
  • VLAN
  • TCP/IP

Machine Learning & AI

  • LLM Fine-tuning · QLoRA
  • Mistral 7B
  • MLX
  • HuggingFace
  • TTS · F5-TTS / XTTS-v2

Hardware & Electronics

  • UART / Serial
  • Oscilloscope
  • RF Signal Generation
  • PCB Analysis
  • Server Hardware

Cybersecurity

  • CTF · Web / Pwn / RE
  • Passive Network Analysis
  • Wireshark
  • Forensics
  • GDPR

Languages

  • Python
  • Swift / SwiftUI
  • Bash
  • JavaScript

Engagements

  • UVS Kemister · Board Member
  • UVS Ingenjörer · Nominations
  • Boo Gårds School

§· Trajectory

Two years, from first rack to reactor physics.

  1. 2024 → '26

    The rack begins

    Cold outreach to 34 companies; Conapto and GleSYS come through by winter '25/26, and eighteen enterprise servers land the spring after — the 42U rack fills.

  2. 2025 · early

    Into security

    Passive 802.11 analysis surfaces a municipal L2 broadcast domain. Wireshark, no active probing.

  3. 2026 · spring

    SSM national CTF final

    14th of 22 as a two-person team (three allowed), and among the youngest — the field averaging 3–4 years older. Web, pwn, forensics, RE. Halted the organiser's screen recording on GDPR grounds.

  4. 2026 · spring

    A device ban, declared void

    Sole complainant; the school-wide ban formally overturned on the Education Act and proportionality.

  5. 2026 · summer

    Hands on firmware

    Scope-profiled a STIHL iMOW PCB, found live UART pads, opened a serial console into its embedded Linux.

  6. 2026 · autumn

    LAIW — the model learns law

    ~133k Swedish/EU legal docs preprocessed and QLoRA-fine-tuned; a live retrieval layer at inference then grounds the citations — hallucinations 0/5 → 5/5 in a small public PoC.

  7. 2026 · now

    Reactor physics & on-device AI

    A real-time PWR/BWR/RBMK simulator, and VisionAI running gesture + gaze entirely on the Neural Engine.

7B weights · 4-bit · on-device

§06 · Contact

Available for research, collaboration, and hard hardware problems.

A. WESSBLAD HÅRDH INITIALISING
0%